|Re: [eigen] Adding zeroes to sparse matrix alters minimum degree ordering|
[ Thread Index |
| More lists.tuxfamily.org/eigen Archives
- To: eigen@xxxxxxxxxxxxxxxxxxx
- Subject: Re: [eigen] Adding zeroes to sparse matrix alters minimum degree ordering
- From: Cedric Doucet <cedric.doucet@xxxxxxxx>
- Date: Wed, 2 Mar 2016 21:01:53 +0100 (CET)
- Thread-index: XoAHGmTcaLrleppZuPHd5oVK+1uCMw==
- Thread-topic: Adding zeroes to sparse matrix alters minimum degree ordering
as far as I remember, minimum degree ordering algorithms are not based on values but on locations of nonzeros.
Therefore, when you add a zero as a "symbolic" nonzero entry of your matrix, you modify de facto the degree of nodes of the graph corresponding to your sparse matrix.
That's the reason why it is not surprising the two permutations differ in your example.
De: "Avi Robinson-Mosher" <avi.mosher@xxxxxxxxx>
Envoyé: Mercredi 2 Mars 2016 20:57:51
Objet: [eigen] Adding zeroes to sparse matrix alters minimum degree ordering
I'm encountering an issue with Incomplete Cholesky. I have a case where the factorization is successful for a particular matrix, but when I add zeroes to the matrix (that is, set entries in the matrix that were implicit zeroes to explicit zeroes) the factorization fails. I don't completely understand why this happens, but I have observed that the permutation produced by AMDOrdering is different for the two cases (and that's a level of magic too dark for me to dig into directly).
I haven't produced a minimum working example (I'm only observing this on medium-sized systems at the moment, 83x83), but I will attempt to do so if necessary. I'm hoping that someone may have an insight without it though..